C4 - AutoML

• Automated machine learning (AutoML) is the process of automating the time-consuming, iterative tasks of building, tuning, and deploying machine learning models

• eg: automated exploration of best approach: logistic regression, random forest, gradient boosting, etc

• the problem must still be defined, and information such as task, training data, target column, primary metric, compute, limits, etc must still be provided

• then AutoML automates much of the candidate exploration

• it optimises according to the selected primary metric

• sweep jobs are still better when the training algorithm has been chosen, and hyperparameter search needs to be done

• they both involve multiple trails, the difference being who controls the search space and at what abstraction level

• AutoML can automate aspects of featurization/pre-processing for supported scenarios

automate aspects of **featurization/preprocessing** for supported scenarios

• automated exploration could become expensive without limits and constrains, so configure constraints such as:

  • experiment timeout
  • trial time out
  • maximum trials/configurations
  • concurrency

• early stopping/termination can also help

• AutoML outputs need tracking